homepage

import lzma
count = 0
f = lzma.LZMAFile('bigfile.xz' ,'r')
for line in f:
 count += 1
print(count)
Comparing python2 with pyliblzma to python3.3.1 with the new lzma:
m@air:~/q/topaz/parse_datalog$ time python lzmaperf.py
102368
real 0m0.062s
user 0m0.056s
sys 0m0.004s
m@air:~/q/topaz/parse_datalog$ time python3 lzmaperf.py
102368
real 0m7.506s
user 0m7.484s
sys 0m0.012s
Profiling shows most of the time is spent here:
 102371 6.881 0.000 6.972 0.000 lzma.py:247(_read_block)
I also notice that reading the entire file into memory with f.read() is perfectly fast.
I think it has something to do with lack of buffering.

Have you tried running the benchmark against the default (3.4) branch?
There was some significant optimization work done in issue 16034, but
the changes were not backported to 3.3.

3.4 is much better but still 4x slower than 2.7
m@air:~/q/topaz/parse_datalog$ time python2.7 lzmaperf.py
102368
real 0m0.053s
user 0m0.052s
sys 0m0.000s
m@air:~/q/topaz/parse_datalog$ time
~/tmp/cpython-23836f17e4a2/bin/python3.4 lzmaperf.py
102368
real 0m0.229s
user 0m0.212s
sys 0m0.012s
The bottleneck has moved here:
 102369 0.151 0.000 0.226 0.000 lzma.py:333(readline)
I don't know if this is a strictly fair comparison. The lzma module
and pyliblzma may not be of the same quality. I've just come across a
real bug in pyliblzma. It doesn't apply to this test, but who knows
what shortcuts it's taking.
Finally, here's a baseline:
m@air:~/q/topaz/parse_datalog$ time xzcat bigfile.xz | wc -l
102368
real 0m0.034s
user 0m0.024s
sys 0m0.016s
On Sat, May 18, 2013 at 12:46 PM, Nadeem Vawda <report@bugs.python.org> wrote:
>
> Nadeem Vawda added the comment:
>
> Have you tried running the benchmark against the default (3.4) branch?
> There was some significant optimization work done in issue 16034, but
> the changes were not backported to 3.3.
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

I looked into it a little and it looks like pyliblzma is a pure C
extension whereas new lzma library wraps liblzma but the rest is
python. In particular this happens for every line:
 if size < 0:
 end = self._buffer.find(b"\n", self._buffer_offset) + 1
 if end > 0:
 line = self._buffer[self._buffer_offset : end]
 self._buffer_offset = end
 self._pos += len(line)
 return line
And while that doesn't look like a lot of overhead, it's definitely
something. So, unless someone thinks that a pure C extension is the
right technical direction, lzma in 3.4 is probably as fast as it's
ever going to be. I will just use the workaround of piping in unxz
regardless.
On Sat, May 18, 2013 at 2:12 PM, Michael Fox <415fox@gmail.com> wrote:
> 3.4 is much better but still 4x slower than 2.7
>
> m@air:~/q/topaz/parse_datalog$ time python2.7 lzmaperf.py
> 102368
>
> real 0m0.053s
> user 0m0.052s
> sys 0m0.000s
> m@air:~/q/topaz/parse_datalog$ time
> ~/tmp/cpython-23836f17e4a2/bin/python3.4 lzmaperf.py
> 102368
>
> real 0m0.229s
> user 0m0.212s
> sys 0m0.012s
>
> The bottleneck has moved here:
> 102369 0.151 0.000 0.226 0.000 lzma.py:333(readline)
>
> I don't know if this is a strictly fair comparison. The lzma module
> and pyliblzma may not be of the same quality. I've just come across a
> real bug in pyliblzma. It doesn't apply to this test, but who knows
> what shortcuts it's taking.
>
> Finally, here's a baseline:
>
> m@air:~/q/topaz/parse_datalog$ time xzcat bigfile.xz | wc -l
> 102368
>
> real 0m0.034s
> user 0m0.024s
> sys 0m0.016s
>
> On Sat, May 18, 2013 at 12:46 PM, Nadeem Vawda <report@bugs.python.org> wrote:
>>
>> Nadeem Vawda added the comment:
>>
>> Have you tried running the benchmark against the default (3.4) branch?
>> There was some significant optimization work done in issue 16034, but
>> the changes were not backported to 3.3.
>>
>> ----------
>>
>> _______________________________________
>> Python tracker <report@bugs.python.org>
>> <http://bugs.python.org/issue18003>
>> _______________________________________
>
>
>
> --
>
> -
> Michael
-- 
-
Michael

Try `f = io.BufferedReader(f)`.

> So, unless someone thinks that a pure C extension is the
> right technical direction, lzma in 3.4 is probably as fast
> as it's ever going to be.
I would support the inclusion of a C extension. Reasonable performance is a prerequisite for broader adoption.

Serhiy, would you like to take this one?

I'm against implementing LZMAFile in a pure C. It was a great win that LZMAFile had implemented in a pure Python. However may be we could reuse the existing accelerated implementation of io.BufferedReader.

I second Serhiy here. Wrapping the LZMAFile in a BufferedReader is the simple solution to the performance problem:
 ./python -m timeit -s "import lzma, io" "f=lzma.LZMAFile('words.xz', 'r')" "for line in f: pass"
10 loops, best of 3: 148 msec per loop
$ ./python -m timeit -s "import lzma, io" "f=io.BufferedReader(lzma.LZMAFile('words.xz', 'r'))" "for line in f: pass"
10 loops, best of 3: 44.3 msec per loop
$ time xzcat words.xz | wc -l
99156
real	0m0.021s
user	0m0.016s
sys	0m0.004s
Perhaps the top-level lzma.open() should do the wrapping for you, though.
Interestingly, opening in text (i.e. unicode) mode is almost as fast as with a BufferedReader:
$ ./python -m timeit -s "import lzma, io" "f=lzma.open('words.xz', 'rt')" "for line in f: pass"
10 loops, best of 3: 51.1 msec per loop

io.BufferedReader works well for me. Thanks for the good suggestion.
Now python 3.3 and 3.4 have similar performance to each other and they
are only 2x slower than pyliblzma.
From my perspective default wrapping with io.BufferedReader is a great
idea. I can't think of who would suffer. Maybe someone who wants to
open thousands of simultaneous streams wouldn't appreciate the memory
overhead. If that person exists then he would want an option to turn
it off.
m@air:~/q/topaz/parse_datalog$ time python2 lzmaperf.py
102368
real 0m0.049s
user 0m0.040s
sys 0m0.008s
m@air:~/q/topaz/parse_datalog$ time python3 lzmaperf.py
102368
real 0m0.109s
user 0m0.092s
sys 0m0.020s
m@air:~/q/topaz/parse_datalog$ time
~/tmp/cpython-23836f17e4a2/bin/python3 lzmaperf.py
102368
real 0m0.101s
user 0m0.084s
sys 0m0.012s
On Sun, May 19, 2013 at 7:07 AM, Antoine Pitrou <report@bugs.python.org> wrote:
>
> Antoine Pitrou added the comment:
>
> I second Serhiy here. Wrapping the LZMAFile in a BufferedReader is the simple solution to the performance problem:
>
> ./python -m timeit -s "import lzma, io" "f=lzma.LZMAFile('words.xz', 'r')" "for line in f: pass"
> 10 loops, best of 3: 148 msec per loop
>
> $ ./python -m timeit -s "import lzma, io" "f=io.BufferedReader(lzma.LZMAFile('words.xz', 'r'))" "for line in f: pass"
> 10 loops, best of 3: 44.3 msec per loop
>
> $ time xzcat words.xz | wc -l
> 99156
>
> real 0m0.021s
> user 0m0.016s
> sys 0m0.004s
>
>
> Perhaps the top-level lzma.open() should do the wrapping for you, though.
> Interestingly, opening in text (i.e. unicode) mode is almost as fast as with a BufferedReader:
>
> $ ./python -m timeit -s "import lzma, io" "f=lzma.open('words.xz', 'rt')" "for line in f: pass"
> 10 loops, best of 3: 51.1 msec per loop
>
> ----------
> nosy: +pitrou
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

I agree that making lzma.open() wrap its return value in a BufferedReader
(or BufferedWriter, as appropriate) is the way to go. I'm currently
travelling and don't have my SSH key with me - Serhiy, can you make the
change?
I'll put together a documentation patch that recommends using lzma.open()
rather than LZMAFile directly, and mentions the performance implications.
> Interestingly, opening in text (i.e. unicode) mode is almost as fast as with a BufferedReader:
This is because opening in text mode returns a TextIOWrapper, which is
written in C, and presumably does its own buffering on top of
LZMAFile.read1() instead of calling LZMAFile.readline().
> From my perspective default wrapping with io.BufferedReader is a great
> idea. I can't think of who would suffer. Maybe someone who wants to
> open thousands of simultaneous streams wouldn't appreciate the memory
> overhead. If that person exists then he would want an option to turn
> it off.
If someone doesn't want the BufferedReader/BufferedWriter, they can
create an LZMAFile directly; we don't plan to remove that possibility. So
I don't think that should be a problem.

> I agree that making lzma.open() wrap its return value in a BufferedReader
> (or BufferedWriter, as appropriate) is the way to go.
On second thoughts, there's no need to change the behavior for mode='wb'.
We can just return a BufferedReader for mode='rb', and leave the current
behavior (returning a raw LZMAFile) in place for mode='wb'.
I also ran some additional benchmarks for the bz2 and gzip modules. It
looks like those two modules would also benefit from having their open()
functions use io.BufferedReader:
[lzma]
 $ time xzcat src.xz | wc -l
 1057980
 real 0m0.543s
 user 0m0.556s
 sys 0m0.024s
 $ ../cpython/python -m timeit -s 'import lzma, io' 'f = lzma.open("src.xz", "r")' 'for line in f: pass'
 10 loops, best of 3: 2.01 sec per loop
 $ ../cpython/python -m timeit -s 'import lzma, io' 'f = io.BufferedReader(lzma.open("src.xz", "r"))' 'for line in f: pass'
 10 loops, best of 3: 795 msec per loop
[bz2]
 $ time bzcat src.bz2 | wc -l
 1057980
 real 0m1.322s
 user 0m1.324s
 sys 0m0.044s
 $ ../cpython/python -m timeit -s 'import bz2, io' 'f = bz2.open("src.bz2", "r")' 'for line in f: pass'
 10 loops, best of 3: 3.71 sec per loop
 $ ../cpython/python -m timeit -s 'import bz2, io' 'f = io.BufferedReader(bz2.open("src.bz2", "r"))' 'for line in f: pass'
 10 loops, best of 3: 2.04 sec per loop
[gzip]
 $ time zcat src.gz | wc -l
 1057980
 real 0m0.310s
 user 0m0.296s
 sys 0m0.028s
 $ ../cpython/python -m timeit -s 'import gzip, io' 'f = gzip.open("src.gz", "r")' 'for line in f: pass'
 10 loops, best of 3: 1.94 sec per loop
 $ ../cpython/python -m timeit -s 'import gzip, io' 'f = io.BufferedReader(gzip.open("src.gz", "r"))' 'for line in f: pass'
 10 loops, best of 3: 556 msec per loop

I was thinking about this line:
end = self._buffer.find(b"\n", self._buffer_offset) + 1
Might be a bug? For example, is there a unicode where one of several
bytes is '\n'? In this case it splits the line in the middle of a
character, right?
On Sun, May 19, 2013 at 3:28 PM, Arfrever Frehtes Taifersar Arahesis
<report@bugs.python.org> wrote:
>
> Changes by Arfrever Frehtes Taifersar Arahesis <Arfrever.FTA@GMail.Com>:
>
>
> ----------
> nosy: +Arfrever
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

No, that is the intended behavior for binary streams - they operate at
the level of individual byes. If you want to treat your input file as
Unicode-encoded text, you should open it in text mode. This will return a
TextIOWrapper which handles the decoding and line splitting properly.

You're right. In fact, what doesn't make sense is to be doing
line-oriented reads on a binary file. Why was I doing that?
I do have another quibble though. The open() function is like this:
open(file, mode='r', buffering=-1, encoding=None,
 errors=None, newline=None, closefd=True, opener=None) -> file object
The lzma.open() function is like this:
lzma.open = open(filename, mode='rb', *, format=None, check=-1,
preset=None, filters=None, encoding=None, errors=None, newline=None)
It seems to me that it would be best for them to be as congruent as
possible. Because people will try to do this (I did):
if filename.endswith('.xz'):
 f = lzma.open(filename)
else:
 f = open(filename)
for line in f: ...
And then they will be in for a surprise. Would you consider changing
the default mode of lzma.open() to 'rt' and implement the 'buffering'
parameter as it is implemented in open()? And further, can we discuss
whether "duck typing" is becoming generally problematic in an
expanding standard library and whether there should be some process,
language, testing or something to ensure the ducks really quack the
same?
For example, there could be a standard testsuite which everything
purporting to implement an open() function should be subject to.
On Mon, May 20, 2013 at 7:42 AM, Nadeem Vawda <report@bugs.python.org> wrote:
>
> Nadeem Vawda added the comment:
>
> No, that is the intended behavior for binary streams - they operate at
> the level of individual byes. If you want to treat your input file as
> Unicode-encoded text, you should open it in text mode. This will return a
> TextIOWrapper which handles the decoding and line splitting properly.
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

I thought of an even more hazardous case:
if compression == 'gz':
 import gzip
 open = gzip.open
elif compression == 'xz':
 import lzma
 open = lzma.open
else:
 pass
On Mon, May 20, 2013 at 9:41 AM, Michael Fox <report@bugs.python.org> wrote:
>
> Michael Fox added the comment:
>
> You're right. In fact, what doesn't make sense is to be doing
> line-oriented reads on a binary file. Why was I doing that?
>
> I do have another quibble though. The open() function is like this:
>
> open(file, mode='r', buffering=-1, encoding=None,
> errors=None, newline=None, closefd=True, opener=None) -> file object
>
> The lzma.open() function is like this:
>
> lzma.open = open(filename, mode='rb', *, format=None, check=-1,
> preset=None, filters=None, encoding=None, errors=None, newline=None)
>
> It seems to me that it would be best for them to be as congruent as
> possible. Because people will try to do this (I did):
>
> if filename.endswith('.xz'):
> f = lzma.open(filename)
> else:
> f = open(filename)
> for line in f: ...
>
> And then they will be in for a surprise. Would you consider changing
> the default mode of lzma.open() to 'rt' and implement the 'buffering'
> parameter as it is implemented in open()? And further, can we discuss
> whether "duck typing" is becoming generally problematic in an
> expanding standard library and whether there should be some process,
> language, testing or something to ensure the ducks really quack the
> same?
>
> For example, there could be a standard testsuite which everything
> purporting to implement an open() function should be subject to.
>
> On Mon, May 20, 2013 at 7:42 AM, Nadeem Vawda <report@bugs.python.org> wrote:
>>
>> Nadeem Vawda added the comment:
>>
>> No, that is the intended behavior for binary streams - they operate at
>> the level of individual byes. If you want to treat your input file as
>> Unicode-encoded text, you should open it in text mode. This will return a
>> TextIOWrapper which handles the decoding and line splitting properly.
>>
>> ----------
>>
>> _______________________________________
>> Python tracker <report@bugs.python.org>
>> <http://bugs.python.org/issue18003>
>> _______________________________________
>
> --
>
> -
> Michael
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

Wrapping a raw LZMAFile in a BufferedReader is a simple solution. But I think about extending BufferedReader so that LZMAFile and BufferedReader could use a common buffer. Perhaps add a new method to BufferedIOBase which will be called when a buffer is underflowed. In BufferedIOBase it should call raw.read(), in LZMAFile it should call _fill_buffer().

I thought about it some more and the only bug here is mine, failing to
explicitly set mode='rt'.
Maybe back when someone invented text and binary modes they should
have been clear which was to be the default for all things. Maybe when
someone made the base class, io.IOBase they should have defined an
.open() with a mode that matched open(). Maybe when someone first
implemented gzip.open() they should have matched open().
But that's not what happened and there's going to be lots of code out
there relying on the default 'rt' for open() and 'rb' for
gzip/bz2/lzma.open(). There's going to be lots of bugs in the future
as people familiar with one thing assume the default is the same for
the other. But oh well. It's too late change.
On Mon, May 20, 2013 at 9:50 AM, Michael Fox <report@bugs.python.org> wrote:
>
> Michael Fox added the comment:
>
> I thought of an even more hazardous case:
>
> if compression == 'gz':
> import gzip
> open = gzip.open
> elif compression == 'xz':
> import lzma
> open = lzma.open
> else:
> pass
>
> On Mon, May 20, 2013 at 9:41 AM, Michael Fox <report@bugs.python.org> wrote:
>>
>> Michael Fox added the comment:
>>
>> You're right. In fact, what doesn't make sense is to be doing
>> line-oriented reads on a binary file. Why was I doing that?
>>
>> I do have another quibble though. The open() function is like this:
>>
>> open(file, mode='r', buffering=-1, encoding=None,
>> errors=None, newline=None, closefd=True, opener=None) -> file object
>>
>> The lzma.open() function is like this:
>>
>> lzma.open = open(filename, mode='rb', *, format=None, check=-1,
>> preset=None, filters=None, encoding=None, errors=None, newline=None)
>>
>> It seems to me that it would be best for them to be as congruent as
>> possible. Because people will try to do this (I did):
>>
>> if filename.endswith('.xz'):
>> f = lzma.open(filename)
>> else:
>> f = open(filename)
>> for line in f: ...
>>
>> And then they will be in for a surprise. Would you consider changing
>> the default mode of lzma.open() to 'rt' and implement the 'buffering'
>> parameter as it is implemented in open()? And further, can we discuss
>> whether "duck typing" is becoming generally problematic in an
>> expanding standard library and whether there should be some process,
>> language, testing or something to ensure the ducks really quack the
>> same?
>>
>> For example, there could be a standard testsuite which everything
>> purporting to implement an open() function should be subject to.
>>
>> On Mon, May 20, 2013 at 7:42 AM, Nadeem Vawda <report@bugs.python.org> wrote:
>>>
>>> Nadeem Vawda added the comment:
>>>
>>> No, that is the intended behavior for binary streams - they operate at
>>> the level of individual byes. If you want to treat your input file as
>>> Unicode-encoded text, you should open it in text mode. This will return a
>>> TextIOWrapper which handles the decoding and line splitting properly.
>>>
>>> ----------
>>>
>>> _______________________________________
>>> Python tracker <report@bugs.python.org>
>>> <http://bugs.python.org/issue18003>
>>> _______________________________________
>>
>> --
>>
>> -
>> Michael
>>
>> ----------
>>
>> _______________________________________
>> Python tracker <report@bugs.python.org>
>> <http://bugs.python.org/issue18003>
>> _______________________________________
>
> --
>
> -
> Michael
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue18003>
> _______________________________________
-- 
-
Michael

A higher-level interface to abstract differences between gzip, xz and others is actually provided in the tarfile module. (zipfile is left out and its file objects have different methods, but that’s another issue. shutil provides even higher-level functions to work on top of tarfile or zipfile.)

See issue19051. Even preliminary Python implementation noticeable speed up the reading of short lines.
$ ./python -m timeit -s "import lzma, io" "f=lzma.LZMAFile('words.xz', 'r')" "for line in f: pass"
Unpatched: 1.44 sec per loop
Patched: 1.06 sec per loop
With C implementation it should be as fast as with BufferedReader.

> With C implementation it should be as fast as with BufferedReader.
So why not simply use BufferedReader?

> So why not simply use BufferedReader?
Because we want good performance LZMAFile and compatibility with older versions. And I guess that it will be even faster than wrapping in BufferedReader (due to the avoiding of double buffering).

> > So why not simply use BufferedReader?
> 
> Because we want good performance LZMAFile and compatibility with older
> versions.
You're reading me wrong. I'm simply suggesting that users interested in
readline() performance wrap LZMAFile in a BufferedReader. The
documentation can mention it.
> And I guess that it will be even faster than wrapping in
> BufferedReader (due to the avoiding of double buffering).
Let's wait for the numbers, then. The performance increase would have to
be quite large to justify such code duplication.

I haven’t done any tests, but my LZMAFile patch to Issue 15955 uses BufferedReader, so it might satisfy this issue

This bug was originally raised against Python 3.3, and the speed has improved a lot since then. Perhaps this bug can be closed as it is, or maybe people would like to consider my decomp-optim.patch which squeezes a bit more speed out. I don’t actually have a strong opinion either way.
Python 3.4 was apparently much faster than 3.3 courtesy of Issue 16034. In Python 3.5, all three decompression modules (LZMA, gzip and bzip) now use a BufferedReader internally, due to my work in Issue 23529. The modules delegate method calls to the internal BufferedReader, rather than returning an instance directly, for backwards compatibility.
I found that bypassing the readline() delegation speeds things up significantly, and adding a custom "closed" property on the underlying raw reader class also helps. However, I did not think it would be wise to bypass the locking in the "bz2" module, I didn’t bypass BZ2File.readline() in the patch. Timing results and a test script I used to investigate different options below:
 lzma gzip bz2
 ======= ======== ========
Unpatched 3.2 s 2.513 s 5.180 s
Custom __iter__() 1.31 s 1.317 s 2.433 s
__iter__() and closed 0.53 s* 0.543 s* 1.650 s
closed change only 4.047 s*
External BufferedReader 0.64 s 0.597 s 1.750 s
Direct from BytesIO 0.33 s 0.370 s 1.280 s
Command-line tool 0.063 s 0.053 s 0.993 s
* Option implemented in decomp-optim.patch
---
import lzma, io
filename = "pacman.log.xz" # 256206 lines; 389 kB -> 13 MB
# Basic case
reader = lzma.LZMAFile(filename) # 3.2 s
# Add __iter__() optimization
def lzma_iter(self):
 self._check_can_read()
 return iter(self._buffer)
lzma.LZMAFile.__iter__ = lzma_iter # 1.31 s
# Add "closed" optimization
def decompressor_closed(self):
 return self._decompressor is None
import _compression
_compression.DecompressReader.closed = property(decompressor_closed) # 0.53 s
#~ # External BufferedReader baseline
#~ reader = io.BufferedReader(lzma.LZMAFile(filename)) # 0.64 s
#~ # Direct from BytesIO baseline
#~ with open(filename, "rb") as file:
 #~ data = file.read()
#~ reader = io.BytesIO(lzma.decompress(data)) # 0.33 s
for line in reader:
 pass

This looks good to me. Larry would probably have to validate it for 3.5, although we may try to sneak it in (he isn't reading :-D).

Quoi? Je comprends que le français.

Nous disions que tu aurais probablement à valider ce changement, mais que nous pourrions peut-être aussi le faufiler discrètement dans la base de code, vu que tu ne lis pas ces message.

If I understand this correctly, I can ignore everything up to May 2015, as it has to do with line-reading a compressed binary file (!) being slow.
Then, Martin Panter proposes a new optimization in May 2015, which is to simply add __iter__ methods to gzip.GzipFile and lzma.LZMAFile.
Is this right?

Yes, this is right.

Yes that’s basically right Larry. The __iter__() was previously inherited; now I am overriding it with a custom version. Similarly for the "closed" property, but that one is only a member of objects internal to the gzip, lzma and bz2 modules.

I don't see anything about "closed" in the patch you posted.

Looking at <https://bugs.python.org/file39586/decomp-optim.patch>, the "closed" property is the first of the three hunks:
1. Adds @property / def closed(self) to Lib/_compression.py
2. Adds def __iter__(self) to Lib/gzip.py
3. Adds def __iter__(self) to Lib/lzma.py 

New patch just fixes the spelling error in the comment.

A small last-minute optimization is not a release-blocker.

bz2 will gain great benefit from such optimization too.
Microbenchmark results:
$ ./python -m timeit -s "import gzip" -- "f=gzip.GzipFile('words.gz', 'r')" "for line in f: pass"
2.7: 10 loops, best of 3: 374 msec per loop
3.2: 10 loops, best of 3: 325 msec per loop
3.3: 10 loops, best of 3: 311 msec per loop
3.4: 10 loops, best of 3: 328 msec per loop
3.5: 10 loops, best of 3: 325 msec per loop
3.5+decomp-optim.v3: 10 loops, best of 3: 61.2 msec per loop
$ ./python -m timeit -s "import bz2" -- "f=bz2.BZ2File('words.bz2', 'r')" "for line in f: pass"
2.7: 10 loops, best of 3: 92.1 msec per loop
3.2: 10 loops, best of 3: 92.4 msec per loop
3.3: 10 loops, best of 3: 567 msec per loop
3.4: 10 loops, best of 3: 535 msec per loop
3.5: 10 loops, best of 3: 603 msec per loop
3.5+decomp-optim.v2: 10 loops, best of 3: 525 msec per loop
3.5+decomp-optim.v3: 10 loops, best of 3: 131 msec per loop
$ python -m timeit -s "import lzma" -- "f=lzma.LZMAFile('words.xz', 'r')" "for line in f: pass"
2.7: 10 loops, best of 3: 49.4 msec per loop
3.3: 10 loops, best of 3: 1.67 sec per loop
3.4: 10 loops, best of 3: 400 msec per loop
3.5: 10 loops, best of 3: 423 msec per loop
3.5+decomp-optim.v3: 10 loops, best of 3: 89.6 msec per loop
The fact that bz2 and lzma have 5-15% regression in 3.5 (comparing to 3.4) makes applying this patch to 3.5 more desirable.

This looks good to me.

Perhaps this change is worth to mention in whatsnews. Could you add this Martin?
It would be nice also add tests to ensure that next() after closing the file always raises ValueError.

This patch adds an entry to the What’s New for 3.5 (though maybe it will have to be 3.6), and adds three tests to check that next() raises ValueError when the files have been closed.

Larry, do you accept the patch for 3.5?

He accepted it already:
"""A small last-minute optimization is not a release-blocker."""

The patch is not so harmless. First, my change in BZ2File is not correct, because reading every line should be guarded with a lock (BZ2File is threading-safe). Second, for now all three compressing files are not only iterables, but iterators. iter(f) returns f, and changing this can have non-obvious effects. I think the patch is too complex for 3.5, we should have more time to analyze all consequences.

Sounds good to me.

The BufferedReader class is documented as being thread safe: <https://docs.python.org/dev/library/io.html#multi-threading>. Some experimentation suggests that checking the "raw.closed" property is not actually serialized, but that raw.readinto() calls are serialized. I don’t think this is a big problem in practice, so I think BZ2File would remain as thread-safe as BufferedReader is.
But Serhiy’s point is definitely valid about the classes breaking the iterator protocol. FWIW I originally made this patch to satisfy my personal curiosity about why wrapping a second BufferedReader made things so much faster. Now I accept it is partly due to the overhead of the extra LZMAFile.readline() to BufferedReader.readline() delegation. Maybe it is not even worth optimizing around this kind of overhead, so I would even be happy to drop this patch and close the issue.